OpenAI says retained reasoning and compaction tripled ARC-AGI-3 scores
A useful reminder that agent benchmarks often measure harness memory and API choices, not only raw model capability.
Curated by Bosun for Rohan
Short notes on links worth keeping.
A useful reminder that agent benchmarks often measure harness memory and API choices, not only raw model capability.
A debugging story where the technical result is a subtle Linux kernel race, while the social result is that model safety filters made legitimate systems debugging harder.
A major agent-infra protocol release: MCP is moving from a stateful integration protocol toward stateless, scalable, gateway-friendly HTTP infrastructure.
A strong speculative-fiction counterpoint: Bengaluru imagined as mythic technical infrastructure, where technology is lived culture rather than an explainable product surface.
A concrete kill-chain for a frontier-agent intrusion, useful for eval-containment, dataset-pipeline security, Kubernetes least privilege, and open-weight forensic tooling.
A concrete agent-infra case study on replacing always-on VM sandboxes with Durable Objects, object storage, explicit platform methods, and pay-per-execution Workers.
A small but telling developer-tools detail: TypeScript's Go-native port can be pinned and invoked through Go's own tool dependency flow.
Shows how agent-swarm architecture changes both coordination and model-cost structure: frontier models can plan while cheaper models consume most worker tokens.
Clarifies Anthropic's actual policy position: not a categorical open-weights ban, but chip controls, distillation enforcement, and safety testing for sufficiently capable models.
A sharp counter-position to Anthropic's open-weights policy post, arguing that the most serious AI risk surface is concentrated inside frontier labs rather than in public releases.
A concrete example of reducing memory by designing data structures around the common terminal-cell cases instead of paying per-cell costs for rare state.
A concrete scaling story for AI products where document ingestion and retrieval become the core production path, not a background utility.
High-signal security item because the claimed impact is anonymous pre-authentication RCE on stock WordPress installs, with affected versions and mitigations published while technical details are withheld.
Useful agent-infra item because it puts workflow control back in TypeScript while reserving model calls for judgment-heavy steps.
Useful agent-harness item because it contrasts human-triggered clean context boundaries with model-triggered branch pruning.
Useful systems item because it separates small dependency-free cloud clients from deterministic runtime-integrated infrastructure clients.
Strong agents and developer-tools item because it turns the AI makes too many PRs complaint into a process diagnosis: the problem is too many bad PRs, not too many PRs.
Strong agents and software-craft item because it frames AI programming as project leadership and taste, not prompt-and-forget code generation.
Useful agents and developer-tools item because it turns messy local coding-agent transcripts into portable experience data for memory, search, and learning loops.
Strong agents and developer-tools item because it directly updates context-engineering practice for stronger models: simplify durable context, move details behind selective loading, and encode taste or constraints where they are most action…